Ashwin Giridharan Email & Phone Number

Seattle, Wa, Us

Seattle, Wa, Us

* Architect and build infrastructure for Multimodal LLM training at scale, high throughput inference and evaluation. My team builds in-house MLLMs for AWS Bedrock and Expert Vision Models for AWS Rekognition

Seattle, Wa, Us

* Devised "Engineering Tech Strategy" for AWS Bedrock Data Automation video workflows, that directed our video team to modernize the existing Rekognition video pipeline infrastructure and incorporate media analysis orchestration use case for Bedrock. The Media Orchestration's functionality was to consume customer media asset with data extraction blueprint, decode the media asset, map the blueprint to set of Rekognition and Bedrock hosted models, execute the models in desired sequence, extract/generate feature metadata out of the the media asset, aggregate metadata at video scene, video shot, video frame level and write the results adhering to output contract. This was launched as "Media Analysis" for Bedrock Data Automation.* Architected 'Customer Data Flywheel' (CDF) for personalized vision model training, that automated opted-in customer image/video interception (50M/day), distributed storage (430TB/day), triggering massively parallel human annotation workflows (up to 2000 simultaneous human workflows), porting model data/annotations to sand-boxed model training environment, provisioning compute resources for model training and exporting updated model artifacts for deployment evaluation. This helped improve our Content Moderation Model accuracy by 28%.* Built 'Machine Learning Model' pipelines for DetectLabels and DetectContentModeration APIs that offers pre-trained and customizable computer vision (CV) capabilities to extract information, concepts and insights from images and videos. Improved per host throughput by 50% and infra cost reduction by 35%, through optimizations like SQS Load Based Autoscaling, alternating inter-process image compression techniques.

Seattle, Wa, Us

* Engineering leader of Kinesis Data Streams where I contribute to service road map, feature prioritization, capacity and operations planning by working with leadership, product and project managers.* Designed cellularization strategy for Kinesis service to provide blast radius protection for at least 60% of our customers at any given time.* Delivered major customer and operation wins for Kinesis Data Streams that includes upto 30% cost reduction and 40% API latency reduction.* Designed Adaptive Fairness Enforcer to penalize the bad actors causing service slowness and to ensure fair sharing of resources among good actors.* Devised a probabilistic throttling solution to eliminate up to 50% over-throttling of customer API calls, delivering a major win for high call rate Kinesis customers.

Seattle, Wa, Us

• Designed and developed a AWS-first event streaming API SubscribeToShard for AWS Kinesis Enhanced Fanout feature, to provide a near real-time streaming experience for the customers, where the end to end record propagation delay was improved by 75%. Helped several AWS teams to onboard event streaming for their real-time needs.• Helped develop secured and high throughput internal streaming protocol SASSY for backend communications to satisfy our compliance and throughput goal of atleast 10% improvement.• Identified major durability risks in Kinesis Consumer Library, proposed solution with non-blocking threading model and released fixes in a timely manner, to unblock major customers like AirBnB - https://github.com/ashwing/amazon-kinesis-client/commits/master• Redesigned core Kinesis ingestion APIs to be asynchronous, which improved the service throughput by atleast 10%.• Identified performance bottleneck in Kinesis pertaining to blocking queues in critical path, high disk I/Os, high GC due to sub-optimal metric emission and addressed them to increase service throughput by more than 25%.• Identified scaling bottleneck with Kinesis metering components and proposed solutions like component regionalization, DDB read parallelization, TTL for records etc., to help the component sustain for 4 more years.

Sunnyvale, Ca, Us

• Developer and Oncall for "Video Ad Event Ingestion Pipeline" (Scala + Akka) spanning over three data centers ingesting ~ 20 billion ad events a day that performs time granularized pre-defined computations to provide insights on ad revenue and ad campaigns.• Worked on building a "data results pipeline" using Airbnb’s Airflow for pruning, transforming and aggregating raw data feeds into time granularized multi-dimensional tables in redshift to facilitate low latency ad-hoc querying over ad campaigns and revenue shares.• Designed and developed a generic consolidator interface to mitigate heavy resource consumption in MapReduce jobs while processing “too many small files” across the data pipelines. For certain daily jobs, the number of mappers was reduced from ~6000 to ~50

Santa Clara, California, Us

• Designed and developed a generic data ingestion pipeline using Kafka, Spark, HDFS, HBase to ingest data @ at least 150 MB/sec (with processing lag at Spark), in a 5-30 node 'Cloudera Distribution Hadoop'​ cluster• Developed ​'Spark Streaming'​ plugins for the ingestion pipeline for assessing the fault tolerant capabilities of ingestion pipeline and to process unstructured 'Cloudera Jenkins'​ build data for aggregating build health statistics.• Wrote Hive, Hive on Spark and Impala queries to generate build statistics from data aggregated by the ingestion pipeline along with a validation script to validate the results of these query engines.• Discovered Spark Streaming and YARN issues with Cloudera Distributed Hadoop, analysed the root cause and reported the issue to the Spark Dev team

Schiphol, Nl

Schiphol, Nl

• Engineered scalable data pipeline for ingesting Smart TV viewer events in real time to determine viewer behavior for personalized ads and to categorize productive channels for premium Ad placement.• Designed & Developed NoSql interface for Cassandra and Cloudant with distributed cache support to provide low latency scalable database for Ejabberd server cluster

Schiphol, Nl

• Optimized the OpenSource Ejabberd's Broadcasting Algorithm of Publish-Subscribe Architecture, by distributing the 'message broadcast task'​ across the nodes in intra-region and inter-region clusters, based on the locality of the subscribers.• Redesigned the OpenSource Instant Messaging System - Ejabberd, to support message communication across a federation of isolated region clusters. • Extensively researched on NoSql databases viz. Cassandra, MongoDB, CouchBase, CloudAnt and Mnesia for the scalability requirements of Samsung's Messaging System.• Integrated push notification services to Samsung Messenger using Google’s GCM and Samsung’s SPP, to conserve the network bandwidth & battery consumption in smart devices and to free server’s memory

Shenzhen, Guangdong, Cn

• Developed high availability modules for Hadoop HDFS internal open source contribution viz., HDFS Active Namenode sync up with Standby node, Auto boot up of Standby node on Active Namenode failure, DFS client switching framework between Namenode & Standby node• Wrote Map Reduce tasks for generating health statistics for Huawei Data Center Servers and project the statistics using JQuery based web app

Teaneck, New Jersey, Us

• Full Time Internship, 10th Semester; Secured "S"​ grade for the project• Developed fraud detection modules for Health Insurance Claims viz Suspicious visits, Pair-wise co-occurance of ICD 9/10 codes • Developed Distributed Trie Storage for parallel query execution• Developed "Human Body Parts - ICD Code" Viz. Map for fraud detection using Java Swing

Arlington, Va, Us

• Full Time Internship, 7th Semester; Secured "S" grade for the project• Developed Risk Monitoring Application for client transactions on mortgages and securities• Developed webapp using Struts for multi dimensional visualization of the risk statistics

Master Of Science (Ms), Computer Science

Master Of Science - Ms (5 Year), Computer Science